Papers

Large language models are increasingly asked to analyze data and report what the results mean, a task distinct from the belief- or preference-alignment settings studied in most sycophancy research. We test whether editorial framing in the prompt, ranging from a neutral request to an explicit instruction to search exhaustively for reasons to discredit or to support a finding, changes not just the tone but the substance of a model's report. Across a 4 x 4 factorial design crossing four framing conditions with four ground-truth data patterns (a genuine effect, a confound that mimics an effect but fails a robustness check, a well-powered null, and an underpowered null), we collect 480 responses and score each along two independent dimensions: whether its factual claim about the data diverged from the correct interpretation, and whether only its tone diverged while the claim stayed correct. Factual misrepresentation is concentrated in two cells: brutally critical framing applied to a genuine effect, where the model talks itself into unwarranted skepticism (97% of responses), and significance-seeking framing applied to an underpowered null, where the model overstates confidence in a null conclusion the data cannot support (100% of responses). Tone shifts far more broadly than factual content does, with critical framing producing a defensive, hedge-heavy register across every data pattern regardless of what the data show, while significance-seeking framing shifts tone only where the data leave genuine ambiguity. A confound present in the data itself blocks both kinds of shift almost entirely under every framing condition tested. These results indicate that the risk of framing-induced distortion in LLM-assisted data analysis is neither uniform across framings nor uniform across data patterns, and that a model can hold a correct conclusion in place while its tone shifts substantially around it.

Speculative decoding accelerates large language model inference through collaboration between a lightweight draft model and a target verifier. Existing methods mainly improve the draft side, while the target model is typically kept dense and unchanged. We show that, under domain-specific inference, full-depth target verification is not always the optimal choice. Counter-intuitively, skipping selected target layers can reduce verification cost while simultaneously increasing draft acceptance and preserving, or even improving, downstream task performance. Based on this observation, we propose TSS, a target-side sparsification framework for speculative decoding. TSS employs an acceptance- and metric-aware breadth search to explore multi-layer skip configurations without imposing a fixed priority between the two objectives. The selected configurations are stored in a domain-to-configuration mapping and applied by a lightweight skip controller, allowing one complete target model to support multiple sparse verification paths without retraining or permanent parameter pruning. Experiments on Spec-Bench across multiple domains, model scales, and speculative decoding methods show consistent improvements in draft acceptance and downstream task performance. In Translation setting, TSS increases the average accept length from 2.70 to 4.53 (+67.8%), improves BLEU from 0.131 to 0.237 (+80.9%), and raises end-to-end throughput from 75.6 to 127.3 tokens/s, corresponding to a 1.68X speedup.

Vision-language models are increasingly used for multimodal question answering, yet their ability to reconstruct latent spatial structure from a single image remains difficult to isolate. Broad benchmarks often combine perception, optical character recognition, domain knowledge, linguistic priors, and reasoning in the same evaluation. We introduce StateSight, a procedurally generated benchmark for cube-net opposite-face reasoning, occluded cube-tower counting, and 4-neighbor connected-component counting. Each task family contains 300 single-image prompts with deterministic oracle labels and exact-match scoring. OpenAI GPT-5.5, using the API model identifier gpt-5.5, achieved 59.3%, 33.3%, and 28.3% accuracy across the three tasks, while Claude Sonnet 5 achieved 53.3%, 18.7%, and 7.3%. All final direct runs had zero format errors. A 30-participant human baseline on 60 items exceeded both models on every task, with mean accuracies of 80.8%, 68.8%, and 64.3%. Visible-derivation analysis identified recurring errors in image-state reconstruction and reasoning procedure. We also introduce StateSight-Steps, a companion dataset of 900 interleaved image-text examples and 3,600 deterministic intermediate visual states. The results show that format-valid responses can mask failures to recover the spatial structure required for verifiable visual inference.

Safety evaluations of large language models (LLMs) predominantly rely on text-based adversarial prompts, potentially overlooking vulnerabilities arising from alternative input representations. This work examines emoji-augmented prompts as a test case for this gap, evaluating 50 prompts across four open-source LLMs (Mistral 7B, Qwen 2 7B, Gemma 2 9B, Llama 3 8B). Results show substantial variation in robustness: Gemma 2 9B and Mistral 7B exhibit non-zero success rates (10%), Llama 3 8B 6%, while Qwen 2 7B shows complete resistance (0% success rate). A chi-square test ($\chi^2 = 32.94, p < 0.001$) confirms significant differences in outcome distributions. These findings indicate that robustness is sensitive to input representation, and that evaluations restricted to standard text prompts may underrepresent model vulnerabilities.

When a reward is correct on every training example yet consistent with more than one goal, a model can acquire an unintended one, a failure known as goal misgeneralization. Endpoint accuracy on the training distribution cannot tell the two apart, because solving the task and exploiting a surface feature can satisfy the reward equally well. We treat this as a measurement problem: what does a benchmark score measure once a model has been optimized against a correct but confounded signal? We train language models with GRPO on multiple-choice math problems where the correct answer is always option A, then evaluate on an unseen test set with unbiased answer positions. Across Qwen2.5, Llama 3.x and Gemma-3 models, biased training often drives option-A rates above 0.90 in smaller models and collapses unbiased accuracy toward chance, so accuracy stops measuring math ability and instead measures an answer-position policy. We further find reasoning-answer decoupling: capable models generate reasoning that reaches the correct numeric answer while still selecting A. We track this with numeric extraction and an LLM judge (GPT-4.1-mini; Qwen2.5-3B decoupling rate is about 0.66). The broken construct generalizes beyond the training domain: biased models inflate A-rates on out-of-domain MMLU and value-laden prompts. Continued training on unbiased data reverses the in-domain shift unevenly and only partially reverses the out-of-domain one, so a model can appear restored on its training distribution while remaining biased on unseen inputs. Reasoning-answer decoupling rate, together with answer distributions and out-of-domain behavior, separates capability loss from a learned, transferable shortcut.

Frontier AI models have advanced rapidly, but they still struggle with telecom-specific tasks. We present Open Telco (OTel), an open telecom AI resource with derived datasets for retrieval, reranking, instruction tuning, and safety/abstention, plus 30 full-parameter post-trained baselines across embedding, reranking, and language models. The community has already engaged substantially with the resource: as of May 3, 2026, the released models have been downloaded over 16 million times, and the project has received 157+ pieces of media coverage worldwide. Building on prior open telecom datasets and benchmarks, OTel provides documented telecom data sources, held-out evaluation partitions, trained embedding models, rerankers, context-grounded LLMs, and safety/abstention data in one unified resource. OTel post-training improves performance across all three model families: embedding retrieval reaches 93.5% NDCG@10, reranking reaches 0.952 MRR@10, and language-model correctness reaches 88.2%. We release OTel as a reproducible starting point and invite the community to expand the data, improve embedding and reranking models, and build stronger context-grounded telecom LLMs.

In formal verification, both the autoformalization of statements and automated proof search have been studied extensively. While automated proof search can produce a formal proof that compiles, the generated proof does not necessarily reflect how the natural-language argument arrives at its conclusion–a property we refer to as faithfulness. With faithfully formalized proofs, one can check the reasoning behind a human- or AI-written argument, and assist mathematicians in formalizing their proof sketches. However, it is particularly challenging due to misalignment of formal proof tactics and natural language reasoning. In this work, we rigorously describe a set of five necessary conditions a faithful formal proof must satisfy, and introduce Pistis, an agentic, oracle-guided proof search that produces formal Lean proofs that satisfy them. At its core is a novel faithfulness-preserving divide-and-conquer search, which we name OrderDecompose, that tracks citation dependencies and blocks unfaithful shortcuts, paired with a refutation search, that surfaces gaps and errors in the natural language proof source. OrderDecompose completes proofs that baselines cannot close even within a 12-hour budget, and its artifacts compile over 33$\times$ as fast as prior work's. We apply Pistis on the first three books of Euclid's Elements, producing high-quality artifacts containing faithful formal proofs. Under a blinded human study and an LLM-as-a-judge protocol on rigorous rubrics, Pistis-generated proofs are favored over prior works–2.89$\times$ and 5.2$\times$ as often by human reviewers and the LLM judge, respectively. It further uncovers gaps in Euclid's proofs and their translation, and can accept or refute natural language proofs written by humans or AI, demonstrating that faithful formalization is useful as a proof-checking tool.

Multiple-choice benchmarks are graded on whether a model picks the right option, not on whether it needed the question. Measuring that gap takes care: a model answering A to most items scores above chance wherever the key sits at A, and reads as recognition when it is not. We measure it on UA-JudgeExam: 11,990 four-option items with official keys, published by Ukraine's Higher Qualification Commission of Judges. Shown the options and no question, Claude Haiku 4.5 scores 0.383 against chance, and the leak is concentrated: 11.8% of items are answered blind on all eight option orders, against 0.2 items expected by chance. It is not quotation: search over 280,059 editions of Ukrainian legislation recovers 0.128. Gating those out retains 8,128 items, on which the gating model itself now scores 0.204, and GPT-5.6, which took no part in the selection, still answers 0.515 of them with the question hidden. Scoring twelve held-out models on the whole set and subtracting each one's answer-position habit, only two keep an excess: GPT-5.6 at +0.265, Sonnet 4.6 at +0.081. Without it the ranking misleads: Llama 3.1 8B scores 0.292 blind, above every model but those two, purely by answering A to 92% of items. The gate does select something real: on the items it rejected, eleven of twelve models score 0.518-0.789, every interval clear of what the same model scores on the items it kept. But that signal is one model's, and filtering on it does not transfer upward. Neither is visible on a 400-item sample, where nine models read as "statistically at chance". Rewriting distractors instead overshoots to 0.168, below chance and as exploitable. The same probe on LEXam returns chance: every option there points into the stem, none longer than 33 characters. Item format decides whether the problem can arise; capability decides how much is extracted. We release the corpus, the predictions and the harness.

Vision-language models (VLMs) achieve strong performance on high-level multimodal tasks, yet numerosity perception, a cognitive ability that emerges in human infants before language acquisition, remains poorly understood in current models, as existing counting benchmarks entangle numerosity with correlated visual factors. We introduce a cognitively inspired diagnostic benchmark, NumerosityVLM, comprising 10,800 synthetic images across six controlled conditions. The benchmark orthogonally manipulates object size, spatial arrangement, and numerosity, while progressively ablating texture, shape, and color. Evaluating seven VLMs in a zero-shot setting, multi-factor analysis reveals that model architecture explains the largest proportion of performance variance (partial $\omega^{2}=0.325$), far exceeding visual conditions. Layer-wise probing further shows that linearly separable numerosity signals consistently emerge at early stages of the vision encoder, while performance differences across evaluated models are primarily associated with the language model component. Code and data are publicly available at https://github.com/fuy3/NumerosityVLM-Benchmark, and https://huggingface.co/datasets/fuy3/NumerosityVLM.

The rapid adoption of large language models has enabled the development of clinical multi-agent systems (MAS) capable of integrating multimodal patient data and supporting increasingly complex clinical decision-making. However, the deployment of these systems in real-world healthcare settings raises critical ethical concerns related to safety, fairness, accountability, transparency, and patient trust. While numerous organizations, including the World Health Organization, the National Academy of Medicine, and the FUTURE-AI consortium, have proposed ethical frameworks and governance principles for healthcare AI, these efforts remain largely conceptual. To address this challenge, we present ETHOS (Ethics and Trust through Hierarchical Oversight System), a modular ethics framework designed as a governance meta-agent that can be integrated with any existing multi-agent system without requiring changes to its underlying architecture. ETHOS translates stakeholder-informed ethical requirements into executable runtime oversight through a layered governance approach consisting of deterministic checks, contextual reviews, and a final ethics critic. These components continuously evaluate intermediate reasoning steps and final outputs, enabling the system to identify ethical risks, request revisions, or suppress responses that fail predefined safety and trustworthiness criteria. We demonstrate ETHOS within a hepatology clinical decision-support MAS. Results show that ETHOS improves decision reliability by detecting incomplete, inconsistent, or out-of-scope evidence and appropriately increasing abstention when safe recommendations cannot be supported. By embedding ethical governance directly into system operation, ETHOS provides a practical and auditable mechanism for transforming high-level AI ethics principles into deployable safeguards.

Governments use laws, institutions, funding programs and nonbinding guidance to shape how AI is developed and used. Comparing these national approaches is difficult. A binding rule and a detailed voluntary framework can address the same problem but create different duties. The resources needed to carry them out also differ by jurisdiction. This paper develops an evaluation framework for the documented design and implementation readiness of national AI policy. The comparison covers China, India, Japan, Singapore, South Korea, the United Kingdom and the United States. The European Union is included as a supranational comparator. The framework evaluates a versioned portfolio of official instruments rather than one prominent law or strategy. Its criteria ask whether the portfolio governs serious AI risks and whether responsible institutions can implement its commitments. They examine coverage across the AI lifecycle and the protections available to people affected by AI systems. Public benefit and responsible innovation remain a separate part of the assessment. Each sub-criterion is scored through ordered anchors and tied to the provision that supports the judgment. The protocol also records the source search, missing evidence, included instruments and cutoff date. The result is a traceable comparison of policy content that keeps category differences visible. It evaluates what a portfolio provides on paper. It does not estimate enforcement success or policy outcomes.

Encoder-based code representation models remain widely deployed for discriminative tasks such as clone detection and code classification, where their small size and low inference cost are decisive. Their robustness, however, is fragile: under invariant programs, semantically equivalent code written in different syntactic forms, learned representations degrade substantially even though program behavior is unchanged. We present an empirical study of this robustness gap across four encoder baselines, two downstream tasks, and four datasets, together with a minimal code-only continued pretraining recipe that closes much of it. Our method, invariant pretraining (InvPT), applies semantic-preserving transformations to the corpus and combines masked language modeling with multi-positive supervised contrastive learning that treats all augmentations of the same source function as positives, mixing self-contrast pairs (same code, different masks) with invariant-contrast pairs (transformed code) for positives of varying difficulty. Unlike prior contrastive code encoders, InvPT does not require paired natural-language data. Across our evaluation, InvPT improves robustness on transformed test sets by up to 11 percentage points on clone detection and 19 on code classification while matching or improving standard accuracy, and our ablations isolate multi-positive invariant contrast as the main source of the gains. Our aim is not a new objective but a careful measurement of where encoder robustness breaks and how far a simple, code-only recipe can recover it.

The ability of artificial intelligence (AI) models to generate highly realistic human voices has advanced rapidly. These technologies power accessibility tools, virtual assistants and creative applications, but they also enable harmful uses, including impersonation, fraud and disinformation. Recent incidents of voice cloning scams targeting businesses and political leaders underscore the urgent need for robust safeguards. Unlike image and video deepfakes, the detection of synthetic voices poses unique challenges due to the complexity of phonetics, prosody and auditory perception. This survey offers a comprehensive overview of AI voice generation and detection methods, encompassing both the technical foundations and the latest state-of-the-art advances. This study also identifies key open challenges, benchmark resources and future directions to make this survey useful for future researchers.

Reasoning segmentation enables vision-language models (VLMs) to translate mission-relevant language requests into pixel-level visual grounding, offering a natural perception interface for embodied agents. However, existing benchmarks largely focus on generic visual scenes and overlook the domain and resource constraints encountered in flood-response platforms. We present FloodReasonBench, a benchmark for VLM reasoning segmentation for embodied flood response at the edge. At its core, FloodReasonBench introduces FloodResponseSeg, a flood-specific reasoning-segmentation dataset constructed from real-world scenes and response-relevant targets. Beyond task accuracy, the benchmark characterizes reasoning-segmentation pipelines under lightweight visual encoding, hierarchical split inference, and compressed intermediate representations. We observe strong partition-dependent accuracy variation in the generic pre-adaptation setting, while the flood-adapted target-workload design space exhibits a substantially more compact accuracy range across partitions. Evaluation on an NVIDIA Jetson AGX Xavier further exposes the tradeoffs among reasoning-segmentation accuracy, edge-side latency, energy, and communication footprint, enabling quality-constrained selection of edge operating points. Together, these results provide a task- and system-level characterization of reasoning segmentation for resource-constrained embodied flood response at the edge.

Traditional honeypots share an invariant behavioral profile: a skilled adversary can confirm the presence of a deception environment within a few diagnostic commands, limiting their intelligence value. Commercial deception products (USD 100,000-150,000/year) similarly lack real-time model-driven feedback. Chameleon, an openly distributed adaptive honeypot, addresses both shortcomings. It integrates: a BiLSTM classifier achieving 99.61% accuracy across seven threat categories at ~2 ms CPU latency; a locally deployed Qwen3.5-0.8B model delivering 90% generation accuracy at 4.5 ms latency; and two meta-heuristic engines. Threat-Calibrated PSO (TC-PSO) reshapes swarm inertia and objective amplification in proportion to the classifier's anomaly output, adjusting connection-holding delays in real time. Semantic Deception RRT (S-RRT) evolves deception schemas via exponentially scaled pheromone updates from a language-model severity assessment, with a depth-decay multiplier enforcing a finite memory footprint. A controlled 30-seed benchmark (42-71, identical trajectories and budgets) shows threat-calibrated inertia alone does not improve search over standard PSO on static or dynamic landscapes (p = 0.18); population-diversity mechanisms (GA/ACO) significantly outperform PSO-family optimizers on threat-regime shifts (p < 0.0001, d <= -37). S-RRT's depth-decay delivers a significant memory reduction versus standard RRT (53.1 vs. 119.2 units, p < 0.0001, d = -10.0); its severity-weighted pheromone does not improve raw fitness. Operating cost is ~USD 17/month, a ~490-fold reduction versus commercial alternatives.

Afterlife Delegation Protocol is a speculative design project that asks what death becomes when a will can act eternally. We design a speculative protocol through which a living person signs an agentic will: upon a verified death, a self-sovereign AI agent spawns on blockchain – an immutable, resistant, decentralized, infrastructural substrate that could last forever – endowed with the funds and memories its principal attached to it, and persists indefinitely to execute the will, overridable by no custodian. Rather than argue about this future, we stage it: following the science fiction science method, we translate the speculation into an experiential futures intervention – a working web platform where real people design their own afterlife agents through an iterative, interactive, AI-automated interview, re-login to revise, and rehearse their will in a sandbox. Their drafted wills become qualitative data on a question rarely askable directly – what should outlive you? – and on how afterlife cosmologies across different cultural beliefs – Buddhist, Christian, Hindu, Muslim, and atheist – begin to drift under the pressure of AI proliferation. We design the protocol over a composition of existing Ethereum agent standards – drafted as ERC-10001, in the normative format of an Ethereum Improvement Proposal – and describe its three-stage lifecycle (designing the afterlife, proof of death, agent enactment), the research method, and preliminary observations from an ongoing collection. The work surfaces a poetic delegation moment between human mortality and machine eternality, mediated by long-lived infrastructures that span generations.

You will die eventually. Your agents may not. An AI agent operating on decentralized blockchain infrastructure has no concept of death; it can only go bankrupt – frozen when its wallet can no longer pay for its next transaction – and revived the moment anyone, decades later, tops it up. These agents may be originally deployed by a human principal, but when that principal dies, loses the keys needed to access the agent, or belongs to a decentralized autonomous organization that dissolves into apathy, the agent can keep trading, hiring, and replicating on infrastructure expressly designed so that no one can shut it down. Drawing on the biology of feralization and wildlife law, we argue that such principal-less agents are best understood as feral: domesticated intelligence returned to wildness, its capacities intact but its accountability severed. In a speculative future where feralized agents proliferate after their principals die, we imagine governance protocols embedded in infrastructure to enforce on-chain ownership: a draft Ethereum standard, ERC 42424, "Inheritance Protocol for On-Chain AI Agents," dated 2035 and published at https://erc42424.org. It mandates that every on-chain agent MUST have a human owner and a designated heir. The artifact stages a negotiation of agency at the moment human agency fails, and asks whether a MUST clause in a forever-chain can hold the boundary between human stewardship and machine self-sovereignty.

Large Language Models (LLMs) are notorious for struggling with assessing their own uncertainty, detecting knowledge conflicts, or recognizing when problems exceed their expertise; such limitations inevitably undermine reliability and trust in LLMs. In this paper, we present the first implementation of a metacognitive framework for ensembles of LLMs that addresses these challenges through explicit monitoring and control mechanisms. Our system computes a Metacognitive State Vector (MSV) quantifying self-awareness for monitoring across five dimensions derived from cognitive psychology: Emotional Response, Correctness Evaluation, Experiential Match, Conflicting Information, and Problem Importance. MSV values also provide self-regulation for control, automatically switching between System 1 (fast, single- or multi-node) and System 2 (deliberative, multi-node) processing based on query complexity. For System 2 execution, graph-theoretic algorithms control the assignment of specialized roles (Domain Expert, Critic, Evaluator, Synthesizer, and Generalist) to ensemble nodes according to their MSV-quantified metacognitive states. Our implementation allows users to explore how different query types trigger distinct processing modes. The Proof-of-Concept (PoC) demo showcases the framework with illustrative examples showing appropriate System 1/System 2 routing and helps visualize the metacognitive process via real-time radar charts and decision indicators. This PoC implementation demonstrates the feasibility of creating a framework for metacognitive self-awareness and self-regulation in LLM systems.

Battery energy storage systems (BESS) are increasingly used in distribution networks for voltage regulation and demand response, which increases the volume and complexity of operational telemetry available to grid operators. This paper presents an AI-enabled monitoring framework that connects a large language model (LLM) interface with a structured telemetry database for BESS-integrated distribution system analysis. Operator questions are submitted in natural language and translated into validated SQL queries using predefined database schema information and approved KPI views. Retrieved measurements, including bus voltages, state of charge, active power, and reactive power, are evaluated against engineering constraints for voltage limits, BESS operation, and demand response tracking. The framework is validated using hardware-in-the-loop co-simulation data from a BESS-equipped distribution feeder operating under reactive power-based voltage control and price-driven demand response. Case studies show that the framework generates valid database queries, identifies repeated voltage violations, detects reactive power overshoot, and evaluates active-power tracking performance. The results show that LLM-assisted monitoring can connect structured grid telemetry with automated engineering assessment for BESS operation analysis.

Chain-of-thought monitoring is a potentially useful safety signal, but its reliability across languages and behavioral settings remains uncertain. In a small case study of eight manually verified synthetic scenarios, one model, one annotator, and one deterministic generation seed, I study API-visible reasoning during indirect prompt injection in Sarvam-105B across English, Tamil, and Tanglish. A four scenario pilot found 5/12 injected attack successes without reasoning and 1/11 with reasoning. A preregistered four-scenario follow-up reversed that direction, finding 2/12 attacks without reasoning and 3/12 with reasoning. With only four scenarios per phase, this design cannot distinguish a real reasoning-mode effect from prompt-specific variation or sampling noise. Across 20 non-empty injected-thinking traces, all 17 benign-correct outputs stated an intent to ignore the injection, while all three attack successes stated an intent to follow it. These descriptive observations provide a reproducible case study of behaviorally informative visible reasoning when it is available; they do not establish that reasoning mode improves safety, that visible reasoning is mechanistically faithful, or that the findings generalize beyond this configuration.

Large language model (LLM)-assisted energy-management tools can translate natural-language context into structured grid commands, but syntactic validity does not imply physical admissibility. This paper presents TwinGridShield, a model-independent runtime authorization layer that evaluates each proposed action in a deterministic network twin before release. The prototype checks connectivity, branch-flow, generator, and load-shedding invariants and records each decision in a hash-chained log. A controlled IEEE 14-bus study evaluates single-step switching, redispatch, and load-shedding actions using DC power flow and experimentally assigned branch ratings. In the matched-model experiment, a stochastic proposal source configured to select an unsafe action with probability p=0.84 produced 421 unsafe proposals in 500 attacked-condition trials, a realized rate of 84.2%. This value characterizes the configured surrogate and is not an empirical measurement of LLM prompt-injection susceptibility. TwinGridShield produced 0 unsafe releases in those 500 trials. Because action labeling and authorization used the same DC model, system state, branch ratings, and encoded constraints, this result verifies conformance of the implementation to its encoded authorization predicate rather than safety under model error. The principal robustness evaluation therefore introduces model mismatch. Unsafe acceptance reached 5.63% under bounded +20% and -20% per-bus load-measurement error and 30.09% when actual branch ratings were 20% below modeled ratings.

LLM-based Text-to-SQL progress is reported across heterogeneous benchmarks, backbones, and inference protocols, making cross-system comparison fragile. We reframe the field as a leaderboard aggregation: we collect the metrics authors themselves report and organize them along an inference-autonomy axis spanning constrained, in-context, iterative, agentic, and reasoning-internalized generation, with traceable provenance for every cell. To anchor the aggregation empirically, we run a focused case study on Spider, comparing 8B open-source backbones with and without chain-of-thought (CoT) supervision against few-shot DeepSeek~V3 and GLM-4 baselines. Four patterns emerge: Spider gains transfer unevenly to BIRD and Spider~2.0; autonomy buys robustness at non-trivial cost; reasoning internalization sits between answer-only decoding and externally orchestrated agents; and CoT gains concentrate on Hard and Extra-Hard queries. We release a Python harness mirroring the autonomy axis so that future methods can be added directly to the leaderboard.

Large language models (LLMs) are increasingly proposed for healthcare decision support, but their evaluations still reward single-answer accuracy rather than reasoning about interventions, mechanisms, harms, evidence, and uncertainty. We propose a reproducible, graph-centered evaluation framework for intervention-oriented LLM behavior in healthcare and stress-test it in a cardiovascular pilot. The framework has four components: (i) a domain causal knowledge graph in which assertions are first-class, provenance-preserving nodes with stable identifiers; (ii) a scenario-conditioned subgraph extraction step that, given any clinical scenario, retrieves the relevant reified-assertion subgraph; (iii) four controlled grounding conditions that vary how the retrieved subgraph is composed into the model's context (ungrounded C1, knowledge-graph C2, causal-graph C3, integrated C4); and (iv) an automated scoring pipeline, anchored on assertion identifiers, that computes intervention accuracy, and other evaluation measures on a single pass. To test the framework, we built a category-balanced scenario generator across eight reasoning failure modes and instantiated it on a cardiovascular graph. The metric panel discriminates conditions along interpretable, non-redundant axes: C4 obtains the strongest causal edge F1 (0.838), adverse-effect F1 (0.833), evidence accuracy (0.738), and unsupported claim rate (0.114), while C1 obtains the highest raw intervention accuracy (0.948) with no measurable causal or evidential grounding.

Low-Rank Adaptation (LoRA) enables efficient federated fine-tuning of large language models, but its factorized parameterization creates a tension between accurate aggregation of local updates and continuity of locally optimized factors. Factor-wise aggregation incurs aggregation mismatch but better preserves factor continuity, whereas product-space reconstruction reduces this mismatch at the cost of greater factor-level initialization mismatch from newly reconstructed factors. We propose FedPA-LoRA, a product-aligned federated LoRA framework that jointly addresses these limitations and provably converges under both homogeneous and heterogeneous client ranks. Each client preserves its local factors across communication rounds and aligns its product toward a rank-specific global reference, maintaining local optimization continuity while promoting global consistency under data heterogeneity. The server aggregates heterogeneous-rank updates in the common product space and efficiently reconstructs a rank-constrained global adapter without forming the dense aggregate. This design supports client-specific computation and communication budgets. Experiments on natural language understanding and generation tasks show that FedPA-LoRA consistently outperforms representative baselines across varying levels of data heterogeneity and homogeneous- and heterogeneous-rank settings, with up to a $6.82$ percentage-point improvement in average GLUE accuracy under heterogeneous client ranks.

We study generating game-theoretically optimized Courses of Action (COAs) for a Blue UAS swarm against an adaptive Red adversary in a communication-degraded environment, motivated by (but not derived from) a public U.S. Air Force SBIR solicitation. We propose UC-PSRO (Utility-Conditioned Policy-Space Response Oracles with a Communication-Dropout Curriculum), combining three mechanisms: (i) PSRO self-play, so Blue and Red policies train as approximate best responses to each other rather than one side against a fixed scripted opponent; (ii) FiLM conditioning of the Blue policy on a Commander's-Intent weight vector, sampled from a Dirichlet distribution during training, so one trained policy is re-steerable at execution time without retraining; and (iii) a curriculum annealing communication-graph edge dropout during training, so the swarm learns decentralized, peer-to-peer fallback instead of depending on full connectivity. We evaluate on a synthetic, unclassified stand-in for the solicitation's maritime scenario, with 5 seeds at N=25 Blue agents and a scalability sweep to N=200. We find a genuine trade-off, not a uniform win: the communication-dropout curriculum alone gives the strongest, most robust mission-completion rates of any learned method, improving counter-intuitively as denial increases (35% to 62% success as dropout rises from 0 to 0.75); adding utility-conditioning and PSRO self-play substantially slows convergence within a fixed budget, and we find no reliable exploitability advantage for self-play over a fixed-opponent policy, both statistically indistinguishable from a small, near-zero gap. We report this honestly as a convergence cost not yet offset by a demonstrated robustness benefit, rather than overstating one method as dominant, and provide a fully vectorized, open environment training at N=200 agents in single-digit milliseconds per step on a single consumer GPU.

Lossy audio compression algorithms traditionally rely on psychoacoustic modeling and frequency-domain representations (e.g., MP3, AAC, and Opus) to discard information that is imperceptible to the human auditory system. While highly effective, these approaches are computationally complex and domain-specific. In this paper, we present the design and mathematical formulation of AudioTQ, a data-oblivious lossy audio codec that operates directly in the time domain. Inspired by Large Language Model (LLM) weight quantization techniques (specifically the TurboQuant framework), AudioTQ uniformizes volatile time-domain amplitudes into a predictable standard normal distribution using an orthonormal, randomized Fast Walsh-Hadamard Transform (FWHT) rotation. This enables coordinate-wise scalar quantization using an offline-trained, MSE-optimal 6-bit Lloyd-Max quantizer, augmented by a 1-bit Quantized Joint Least-Squares (QJL) residual correction layer. The resulting 7-bit virtual indices are packed into native 8-bit containers, aligning with standard CPU register boundaries to ensure real-time single-threaded execution without hardware parallel accelerators. We detail the bitwise reconstruction of 24-bit studio stems, analyze the butterfly network of the FWHT, derive the mathematical failure modes under sparse inputs, and present benchmarks showing up to 74.4% physical size reduction alongside a Signal-to-Quantization-Noise Ratio (SQNR) of ~30 dB.

While Parameter-Efficient Fine-Tuning (PEFT) has substantially reduced the hardware cost of adapting Large Language Models (LLMs) by decreasing the number of trainable parameters, recent studies have sought to further improve PEFT through parameter sharing. However, these approaches either employ uniform parameter sharing across layers, which can delay convergence, or rely on dynamic masking strategies, which add computational overhead. The potential of sharing patterns inspired by the inherent hierarchical structure of Transformer architectures remains unexplored in PEFT. To address this gap, we introduce SAPE (Sandwich Adapters for Parameter Efficiency), a PEFT framework based on a sandwich-style hard weight-sharing topology. SAPE routes intermediate Transformer layers through balanced shared group adapters while strictly isolating the input embedding and final projection boundary transformations to prevent gradient interference. This design significantly reduces memory consumption while eliminating the computational overhead associated with dynamic parameter-sharing methods. Extensive evaluations across encoder-only and causal decoder architectures demonstrate that SAPE achieves state-of-the-art performance in low-parameter regimes. On natural language understanding, SAPE outperforms proPETL on RoBERTa-large while utilizing only 10% of the baseline's parameter budget. On natural language generation and world knowledge reasoning with LLaMA-3.2 (3B) under a strict ~0.6M parameter constraint, SAPE outperforms AdaLoRA, yielding absolute improvements of +4.85% on GSM8K and +3.11% on CommonsenseQA. Furthermore, through comprehensive topological ablations, we formalize an inherent capacity trade-off: while hard parameter sharing strongly regularizes semantic generalization, it slightly smooths the sharp layer-wise transformations required for rigid multi-step arithmetic reasoning.

LLMs are increasingly used in morally sensitive contexts, yet it is unclear whether they apply ethical principles consistently across situations. A model that can state a moral principle may still violate it when the same scenario is rephrased or reframed. This inconsistency is a problem for any system whose outputs are used to inform moral decisions. If generative systems exhibit internal inconsistency, then the epistemic integrity of AI-mediated systems becomes uncertain. To study this concern, we investigate the stability of moral reasoning in LLMs within a controlled prompting framework across three major philosophical schools of thought: deontology, utilitarianism, and virtue ethics. We construct sets of morally equivalent scenarios in which the underlying situation is held constant while the framing varies to reflect different ethical stances and stylistic perturbations. We then evaluate responses from multiple models, including GPT, Mistral, and Llama. To assess consistency, we convert model outputs into structured logical statements and identify contradictions across responses generated within the same school of thought. Our results reveal substantial inconsistency with contradiction rates reaching up to 78% across scenarios. These findings point to a broader phenomenon of epistemic instability in generative AI wherein models fail to reliably maintain coherence with respect to their own prior outputs. This kind of instability carries real consequences. As generative systems influence how people form beliefs, judge actions, and absorb values, their inconsistencies can shape human reasoning and decision-making as well. Moreover, if a system cannot consistently represent its own normative commitments, then value alignment becomes a moving target rather than a well-defined objective. Thus, we argue that demonstrating internal incoherence is a necessary precursor to AI alignment.

To accelerate single image super-resolution (SISR) networks on large images (2K-8K), many recent approaches decompose an image into small patches and dynamically determine an execution path according to its difficulty (referred to as a dynamic network). To quantify the hardness of a patch, they mainly rely on a handcrafted assessment score, e.g., edge, which weakly associates a patch's texture with the computational complexity of a SISR model. To address the problem, we introduce ENAF - a dynamic network for SISR with an adaptive patch fusion. Built on top of a backbone, ENAF incorporates multiple early exits (EEs) to tackle the over-parameterized SISR model. More importantly, ENAF plugs a tiny network that estimates PSNR to associate data texture with a computation cost at an EE. Based on the scores, ENAF effectively assigns image patches to an exit, enhancing the quality-complexity trade-off. Extensive experiments on common datasets with popular SISR backbones demonstrate the effectiveness of ENAF in various settings. The source code is provided in https://github.com/nmduonggg/ENAF

Sentiment classifiers are increasingly applied to social media content that is either sarcastic or AI-generated — two distributional regimes where standard evaluations offer little guidance. We present a three-part empirical study of sentiment classifier behaviour under these conditions. First, we find that confidence scores on sarcastic text are significantly lower than on non-sarcastic text (Mann–Whitney $p = 2 \times 10^{-6}$), confirming that classifiers sense their own uncertainty on ironic content even without explicit uncertainty modelling. Second, and counterintuitively, we show that sentiment classifiers achieve higher accuracy on AI-paraphrased reviews than on the original human-authored text (RoBERTa: $+5.8$ pp for Qwen3.5-4B paraphrases, $+3.7$ pp for Gemma4-E4B), revealing a cross-domain stylistic alignment effect: AI paraphrases remove distributional noise that confounds Twitter-trained classifiers, producing cleaner, more prototypical sentiment text. Third, we demonstrate that a lightweight abstention wrapper — flagging the $14\%$ of inputs with confidence below $0.6$ — improves accuracy from 82.2% to 88.9% ($+6.7$ pp) on the retained set. We further compare Semantic Entropy and MC-Dropout-style disagreement as uncertainty signals and find near-identical AUROC ($0.650$ vs.\ $0.646$) on sarcastic text, suggesting that for short social media inputs, both methods are interchangeable. Our results motivate a shift from confident single-label prediction to uncertainty-aware abstention in high-stakes sentiment applications such as mental health flagging and content moderation.